1. Field of the Invention
The present invention relates to a device and a method for suppressing a noise signal, a device and a method for detecting a special signal, and a device and a method for detecting a notification sound.
2. Description of the Related Art
Methods for suppressing noise signals are widely used in the fields of speech recognition and mobile communications. In the field of speech recognition, noise-signal suppression processing is performed prior to speech recognition processing, thereby suppressing unnecessary noise other than speech signals and improving the accuracy of speech recognition. In the field of mobile communications, noise-signal suppression processing is performed prior to voice coding to prevent the articulation of speech from deteriorating significantly due to mixture of noise into the voice coding at a low bit rate.
Examples of typical noise reduction methods using noise spectrum detection technologies include a spectrum subtraction method. The spectrum subtraction method estimates a noise spectrum and subtracts a noise component from an input signal including noise, thereby reducing the noise.
The spectrum subtraction method recognizes signals having stationary energy, such as white noise and air conditioners, as noise with relative ease, and has a high noise suppression effect. By contrast, the method has difficulty in classifying a signal, which is closer to a tone signal (a sine wave signal) with a high energy component such as a warning sound including a siren sound or a high-pitched engine sound, as a noise signal. These special sounds have frequency components forming a harmonic structure which is a characteristic of speech, and are extremely similar to components of human speech. This makes it difficult to estimate a noise spectrum of the special sounds.
In an analysis of such a noise spectrum of a siren sound, an engine sound, or the like in a short time period (several tens of milliseconds to several hundreds of milliseconds), a harmonic structure having a fundamental frequency and harmonics, which is a characteristic of a speech spectrum, is clearly observed. Because it is difficult to distinguish a noise spectrum from a speech spectrum, a special analysis method is required.
The duration of the analysis described above is substantially equal to the duration of processing of a typical amount of speech signals accumulated in hardware and of various types of signal processing, such as voice coding. The duration of the analysis is appropriate for the duration of a frequency analysis in a field that requires real time processing. To apply the method to a device used for mobile communications that allow only extremely short delay in signal processing, the analysis of a speech signal needs to be finished within the duration of processing described above.
Japanese Patent Application Laid-open No. 2002-258899 discloses a device for suppressing noise that compares a variation period of a frequency spectrum and spectral intensity of noise with a preset pattern, thereby detecting a special noise signal (e.g., a warning sound) serving as a target so as to suppress the noise signal. In the case of a frequency signal being a long periodic signal such as a siren sound whose frequency signal varies with time, the device for suppressing noise extracts a fundamental frequency by a frequency analysis and compares the extracted fundamental frequency information with frequency shift information (a temporal locus of the fundamental frequency) that varies with time, thereby detecting and suppressing a noise signal such as a siren sound.
Japanese Patent Application Laid-open No. 2005-77875 discloses an alarm sound source recognition device that carries out a frequency spectrum analysis of noise, extracts a peak frequency having the maximum value, and determines, when the sharpness of the peak frequency satisfies a predetermined condition, a noise signal serving as a target as an alarm sound. Based on the characteristics of an artificial siren sound and alarm sound that the energy is concentrated in an extremely narrow bandwidth, the alarm sound source recognition device obtains the sharpness of the peak frequency having the maximum value based on an energy ratio with a frequency band adjacent thereto. Thus, the alarm sound source recognition device detects a noise signal such as a siren sound.
U.S. Pat. No. 7,639,147 discloses a detecting device that detects an alarm sound output from a personal alert safety system (PASS). The detecting device analyzes whether the difference between the maximum sound pressure level of collected audio and the minimum sound pressure level thereof exceeds a predetermined reference value. The detecting device further carries out a frequency analysis and a beat pattern analysis in a time domain on the audio, thereby determining whether or not the audio is an alarm sound specific to the PASS. The detecting method is supposed to be able to detect the alarm sound reliably (refer to FIG. 4, FIG. 5, and the description thereof).
The device for suppressing noise disclosed in Japanese Patent Application Laid-open No. 2002-258899 carries out a pattern analysis to detect a warning sound. Because a certain period of time (several seconds) is required to determine the periodicity in a long term, delay occurs until a noise signal is to be suppressed. In addition, the pattern analysis increases the signal processing load.
The detecting method using a pattern analysis basically detects no signal other than a signal registered for the pattern analysis. In the case of a speech signal overlapping with a warning sound signal, the component of the speech signal is likely to be suppressed, resulting in deteriorated articulation of the speech signal.
The alarm sound source recognition device disclosed in Japanese Patent Application Laid-open No. 2005-77875 employs no pattern analysis, thereby placing no restrictions on a detectable alarm sound. Determination of an alarm sound, however, requires extremely sharp energy peak characteristics in a frequency domain. In a relatively quiet environment and an environment with no sound source at a high sound pressure level other than the alarm sound, the sharp energy peak characteristics can be detected with relative ease.
If noise other than an alarm sound is mixed, however, the peak characteristics are blunted by an influence of the noise component. Because the frequency analysis is usually carried out in duration from several tens of milliseconds to several hundreds of milliseconds, characteristics in an extremely short time are detected. As a result, a noise signal that is not related to an alarm sound may possibly have instantaneously sharp peak characteristics. A speech signal particularly has instantaneously sharp peak characteristics and is more likely to be detected mistakenly.
The alarm sound, capable of being detected by the alarm sound source recognition device disclosed in Japanese Patent Application Laid-open No. 2005-77875, is limited to the alarm sound at a fixed frequency having sharp peak characteristics. The alarm sound source recognition device fails to detect a siren sound whose frequency varies in a long period. Ann engine sound originally has no fixed frequency, for example. The alarm sound source recognition device also fails to detect an alarm sound having no sharp peak characteristics because of frequency variations caused by movement of a sound source and movement of a sound collecting unit as a target alarm sound.
The alarm sound detecting device disclosed in U.S. Pat. No. 7,639,147 carries out a pattern analysis to determine whether or not the frequency and the beat pattern of the obtained audio are identical to the frequency and the beat pattern of the PASS, respectively, thereby detecting an alarm sound. The alarm sound detecting device needs to store therein a plurality of patterns in advance, resulting in an increased circuit scale. In addition, identification of the type of the alarm sound extremely increases the analysis time.
In view of the disadvantages described above, it is an object of the present invention to provide a device and a method for detecting a special signal that detect the presence of a special signal in a short time with a smaller amount of memory and operation without being restricted by the type of a warning sound and use conditions and to provide a device and a method for suppressing a noise signal that suppress the detected special signal as a special noise signal.
It is also an object of the present invention to provide a device and a method for detecting a notification sound that detect the presence of a notification sound in a short time with a smaller amount of memory and operation without being restricted by the type of a notification sound and use conditions, and to provide a device and a method for suppressing a noise signal that suppress the detected notification sound as a special noise signal.